AUTO PASS 2 · LOAD

Every phase in this pack's automated pipeline starts at a LOAD node, and SayaImagePhase2Load is the first one you'll hit in sequence - because phase 1 is the only pass with no predecessor. Pass 2 is HIRES / USDU (upscale-denoise-upscale territory), and to do that job it needs the base image that phase 1 produced. This node grabs it.

Here's the design decision that makes these nodes feel weird at first: it has no inputs at all. Zero sockets, no widgets. Instead of reading from wires, it reads from disk - the validated checkpoint that the previous phase's STOP/Review node saved into ComfyUI/output/image/checkpoints/. That's the whole trick of the architecture: phases are isolated runs that communicate through the filesystem, not through connections in one giant graph. Each phase runs as its own prompt, loads its own checkpoint, and hands the baton via an on-disk manifest.

Outputs

Six outputs, identical across all the fixed-phase LOAD nodes:

• image - phase 1's validated image, loaded back as an IMAGE tensor. Your img2img or upscale source.
• manifest_json - the previous phase's full record (seed, prompts, models/VAEs/samplers used, paths).
• checkpoint_path - the validated image's path on disk.
• seed - the seed that made the base image.
• positive_prompt / negative_prompt - what prompted it.

The point of re-exposing seed and prompts: pass 2 can either continue cleanly from them or deliberately drift (new seed, edited prompt), and the manifest says what the base actually was - no guessing two passes later.

When it fails

The most common failure is boring and predictable: phase 1 hasn't validated yet. This node throws if the previous phase's checkpoint doesn't exist or isn't marked validated - the pipeline refuses to upscale an image that was never approved (or that you Restart-discarded). Run the passes in order: let pass 1's STOP (or the Review node's Continue) promote its candidate first. One nice touch: the node watches the previous manifest's file mtime, so if you regenerate phase 1, the cached phase-2 graph automatically notices and picks up the new image on its next run.

Use it only if you're actually running the author's multi-phase pipeline - in a one-shot workflow a plain LoadImage does the same loading with less ceremony. Same install as everything here: ComfyUI Manager search saya-comfy-couple-plus, or clone into custom_nodes and restart. numpy/Pillow only, no downloads. WIP pack - ports can shuffle between versions, so back up workflows before updating.